Heterogeneous Sensing-based In-process Quality Monitoring of Single-tracks Built using Laser Powder Bed Fusion Additive

Manufacturing Process

Aniruddha Gaikwad

Brian Giera, PhD.

Prahalada Rao, PhD.

Solid Freeform Fabrication Symposium, 2019Date: 08/13/2019

Room no. 412Time: 9:15 AM

aniruddha.gaikwad@huskers.unl.eduUniversity of Nebraska-Lincoln

Acknowledgements

National Science Foundation

CMMI 1719388

CMMI 1739696

CMMI 1752069 (CAREER, Smart Additive Manufacturing)

2

Acknowledgements3

The data for this work was generated at Lawrence Livermore National Laboratory (LLNL) by Dr. Brain Giera and colleagues.

Yuan, Bodi, Gabriel M. Guss, Aaron C. Wilson, Stefan P. Hau‐Riege, Phillip J. DePond, Sara McMains, Manyalibo J. Matthews, and Brian Giera. "Machine‐Learning‐Based Monitoring

of Laser Powder Bed Fusion." Advanced Materials Technologies (2018): 1800136.

4

Detect AM part flaws using data from in-situ heterogeneous sensors

Outline

• Introduction

• Objective and Hypothesis

• Experimental Studies

• Methodology and Results

• Conclusions and Future Work

5

Outline

• Introduction

– Background

– Motivation

• Objective and Hypothesis

• Thermal Modeling using Graph Theory

• Experimental Studies and Results

• Conclusion and Future Work

6

Background7

Laser powder bed fusion (LPBF) additive manufacturing (AM) process

Motivation8

Only 2 out of 7 parts were built defect free

Can we build parts without having to print-and-pray?

Part quality inconsistency is major impediment in AM

Outline

• Introduction

• Objective and Hypothesis

• Experimental Studies

• Methodology and Results

• Conclusions and Future Work

9

Objective and Hypothesis10

Develop a cascading artificial neural network (C-ANN) to fuse process signatures acquired from heterogeneous in-situ sensors, and subsequently identify defects

High speed video camera

Pyrometer

Acoustic emission sensor

Sensor Data Fusion

Detect Flaws

Rationale11

Chain a series of neural networks to make step-wise multiple predictions.• Computationally sparse• Can accommodate multiple types of data (image, time series)• Physically interpretable• Multiple Input – Multiple Output

Outline

• Introduction

• Objective and Hypothesis

• Experimental Studies

• Methodology and Results

• Conclusions and Future Work

12

Experimental Setup13

Yuan, Bodi, Gabriel M. Guss, Aaron C. Wilson, Stefan P. Hau‐Riege, Phillip J. DePond, Sara McMains, Manyalibo J. Matthews, and Brian Giera. "Machine‐Learning‐Based Monitoring of Laser Powder Bed Fusion." Advanced Materials Technologies (2018): 1800136.

Heterogeneous in-process sensors coaxial to the laser: video camera and pyrometer

Experimental Data14

Video camera frame rate: 20 kHz• Frame size: 256 × 256 pixels• Video camera is co-aligned with the laser.• Number of frames per video (𝑁): 12 to 50 frames (depends on laser velocity)

Pyrometer sampling rate is 100 kHz (5 times faster than video camera frame rate)

Frame 1

Frame 2

Frame 𝑁

High speed video camera frames Pyrometer data

Laser location

Pyr

om

ete

r re

adin

g

Yuan, Bodi, Gabriel M. Guss, Aaron C. Wilson, Stefan P. Hau‐Riege, Phillip J. DePond, Sara McMains, Manyalibo J. Matthews, and Brian Giera. "Machine‐Learning‐Based Monitoring of Laser Powder Bed Fusion." Advanced Materials Technologies (2018): 1800136.

Experimental Data Used in this Work15

10 video camera frames used per single track

Pyr

om

eter

rea

din

g

Laser location

Pyrometer dataExtracted pyrometer data

Below shown is an example of the data used from the sensors

Experimental Conditions16

• Laser power: 50 W to 375 W – Equal increments of 32.5 W (11 different laser power settings)

• Laser velocity: 100 mm/s to 400 mm/s– Equal increments of 30 mm/s (11 different laser velocity settings)

• Layer thickness: ≈50 µm

• Length of single track is 5 mm

5 m

m

Develop and apply an algorithm to analyze heterogeneous sensor data fusion to monitor quality of single-tracks

Yuan, Bodi, Gabriel M. Guss, Aaron C. Wilson, Stefan P. Hau‐Riege, Phillip J. DePond, Sara McMains, Manyalibo J. Matthews, and Brian Giera. "Machine‐Learning‐Based Monitoring of Laser Powder Bed Fusion." Advanced Materials Technologies (2018): 1800136.

Characterizing Single-Track Quality17

Quality related features are estimated from height maps

5 m

m

Single Tracks deposited using LPBF

Height maps of single-tracks

Characterizing Single-Track Quality18

Mean of width of single track (µ)

Thickness

Standard deviation of width of single track (𝜎)

Wave profile

Continuity of single track (𝛿)

2 m

m

The following single-track quality related features are estimated from the height maps

Outline

• Introduction

• Objective and Hypothesis

• Experimental Studies

• Methodology and Results

– Statistical Feature Extraction from Sensor Data

– Cascading Artificial Neural Network

– Results

• Conclusions and Future Work

19

Statistical Feature Extraction from Sensor Data20

Pyrometer signal →mean, standard deviation, skewness, kurtosis

High speed video camera frames

Meltpool intensity

𝐼𝑎𝑖 =

𝑥=1

𝑀

𝐼𝑥

𝐼𝑎𝑖 = Intensity of meltpool𝐼𝑥= Intensity of a pixel in the meltpool𝑀= number of pixels in the meltpool

Meltpool area

𝐴𝑎𝑖 = 𝜋 × 𝐿𝑚𝑎𝑗𝑜𝑟 × 𝐿𝑚𝑖𝑛𝑜𝑟

𝐴𝑎𝑖 = Area of meltpool

𝐿𝑚𝑎𝑗𝑜𝑟= Major axis length

𝐿𝑚𝑖𝑛𝑜𝑟= Minor axis length

𝑑𝑦

Meltpool circularity

𝑐𝑖𝑟𝑐(𝜇)𝑎𝑖 =

σ𝑦=1𝐸 𝑑𝑦

𝐸

𝑐𝑖𝑟𝑐(𝜎)𝑎𝑖 =

σ𝑦=1𝐸 𝑑𝑦 − 𝑐𝑖𝑟𝑐(𝜇)𝑎

𝑖

𝐸

Cascading Artificial Neural Network (C-ANN)21

Pyrometer

features:

Mean, std.

dev,

kurtosis,

skewness

Global

energy

density

level (𝐸𝑑)

High speed video

camera frame

+

Meltpool features

Network outputNetwork inputFeature set

Shallow

artificial

neural

network I

Mean of

single

track

width (𝜇)

Shallow

artificial

neural

network II

Mean of

single track

width (𝜇)

+𝐸𝑑, high speed video

camera frame features

Shallow

artificial

neural

network III

Standard

dev. of

single track

width (𝜎)

+𝜇, 𝐸𝑑, high

speed video

camera frame

features

Shallow

artificial

neural

network IV

Continuity

of single

track (𝛿)

Below shown is the architecture of the cascading artificial neural network (C-ANN)

22

Training data1620 single-tracks

Validation data180 single-tracks

Testing data298 single-tracks

Ten-fold cross-validation is performed on the Training/Validation data setTesting data is not seen by the trained network

Cascading Artificial Neural Network (C-ANN)

Each printing condition has approximately 17 single-tracks

High Speed Video Camera Features23

Effect of energy density (𝐸𝑑 , 𝐽/𝑚𝑚 ) applied to deposit single-tracks on meltpool area and meltpool intensity

Meltpool area and meltpool intensity increase with increasing 𝐸𝑑

Pyrometer Features24

Below shown is the effect of change in energy density applied to deposit single-tracks on meltpool intensity captured by pyrometer

Meltpool intensity captured by pyrometer increases with increasing 𝐸𝑑

Result: Cascading Artificial Neural Network25

Methodology (sensor)Global energy

density (𝑬𝒅)

Single-track

width (𝝁)

Consistency of single-

track width (𝝈)

Continuity of

single-track

C-ANN (high speed

video camera +

pyrometer)

0.9657 0.8845 0.7726 0.7256

ANN (high speed video

camera)NA 0.7727 0.6571 0.6952

ANN (pyrometer) NA 0.8165 0.7154 0.6525

ANN (high speed video

camera + pyrometer)NA 0.8565 0.7423 0.6884

Deep learning CNN

(high speed video

camera) [1]

NA 0.93 0.73 NA

[1] Yuan, Bodi, Gabriel M. Guss, Aaron C. Wilson, Stefan P. Hau‐Riege, Phillip J. DePond, Sara McMains, Manyalibo J.

Matthews, and Brian Giera. "Machine‐Learning‐Based Monitoring of Laser Powder Bed Fusion." Advanced Materials

Technologies (2018): 1800136.

C-ANN performs better than algorithms with single sensor data

Outline

• Introduction

• Objective and Hypothesis

• Experimental Studies

• Methodology and Results

• Conclusions and Future Work

26

Conclusion and Future Work

• Cascading-artificial neural network (C-ANN) is an approach for sensor data fusion

– Predicts single track quality aspects with statistical fidelity exceeding data from a single sensor.

– Computationally tractable and interpretable compared to deep learning.

• Implement C-ANN for flaw detection, such as porosity and delamination.

27